Research On Off-line Signature Verification

The lack of password technique on security stimulates the research of biometrics. As one of biometrics, off-line signature verification has been widely accepted because of its convenience, efficiency and no invasion of privacy.Currently, the main task of manual handwriting verification is signature verification and its right rate is much higher than manual verification. The differences between automatic signature verification and manual one are feature selecting and the way of judging. In the way of feature selecting, manual verification put more attention to the local features while auto-verification mainly selects global features. In the way of judging, the former judges by the value of the feature while the latter regards features as equality.In this thesis, three factors affected the right rate in all phases of auto-verification have been discussed. The first one is improper image pre-processing, such as segmentation, normalization and thinning, which lose the features of letters relationship and strokes. The second one is improper feature selecting and processing, such as classifying by meaning or the way of extracting, feature normalization. The final one is the judgment, which has no evaluating.Using the manual verification experts'experiences for reference, the writer has put forward proper algorithms for improving image pre-processing, feature selecting and classifier designing. In addition, a voting system based on multi-expert has been proposed.The main innovations of the thesis are as follows.â'ˆAs to pre-processing of signature image, a new normalization algorithm based on mathematical morphology is proposed aimed at questions existing in normal pre-processing, such as segmentation and thinning. It has proved that the differences made by tools can be removed while the details of strokes can be reserved.â'‰As to feature extracting, a novel feature about relationship of letters has been extracted, such as number of connected regions and holes, mean, standard deviation, skewness and kurtosis of the image projection.